The Lift Line
A machine may hand the judge the file, carry the record and translate the argument, but it must never be allowed to decide, because a judgment is an act of reason for which only a human can be held answerable.
Why This Editorial Matters for Your Exam
The Supreme Court of India has circulated a draft framework on the use of generative artificial intelligence in judicial work. Its organising idea is a clean division of labour. AI may perform assistive, non-adjudicatory tasks, transcription of proceedings, translation of orders into regional languages, cause-list scheduling and legal research. It is barred from the core judicial functions, deciding a case, constructing the chain of legal reasoning and drafting the operative judgment. The draft reflects a principle that runs through administrative law and now through technology governance alike: the human-in-the-loop. A decision that affects a citizen’s liberty, property or rights must be authored by an accountable human mind, not generated by a statistical model that cannot be summoned, cross-examined or held to account.
GS Paper 2: the structure, organisation and functioning of the judiciary; e-governance applications and models; and the accountability of institutions exercising public power.
GS Paper 3: developments in science and technology; the governance of artificial intelligence; and issues of data protection and privacy.
For Prelims, hold the specifics: the e-Courts Mission Mode Project, now in its third phase, which digitises court records and case management; SUVAS, the Supreme Court Vidhik Anuvaad Software, which translates judgments into Indian languages; SUPACE, the Supreme Court Portal for Assistance in Court Efficiency, an AI tool for legal research and case briefing that was explicitly designed not to take decisions; the Digital Personal Data Protection Act, 2023, which governs the processing of personal data; and the distinction between an assistive tool and an automated decision system. For Mains, argue that the value of AI in the justice system lies in reducing delay and widening access, while the risks of algorithmic bias, hallucinated case law and opaque reasoning make it unfit for adjudication itself.
Background and Context
Indian courts carry one of the largest case backlogs in the world, with crores of matters pending across the district judiciary, the High Courts and the Supreme Court. Much of the delay is not judicial thinking but judicial paperwork: transcription of oral evidence, translation of orders for litigants who do not read English, scheduling, indexing and the retrieval of precedent. These are exactly the tasks at which large language models excel, and the e-Courts project has spent a decade building the digital rails on which such tools can run. SUVAS already translates judgments into regional languages, and SUPACE was built to brief judges on the facts and the law without ever presuming to advise them on the outcome.
The temptation, once a model can summarise a case file, is to let it summarise the conclusion. That is the line the draft framework refuses to cross, and rightly so. A generative model does not know the law; it predicts the next plausible token. It can produce fluent, confident prose that cites cases which do not exist, a failure already recorded in courts abroad where lawyers filed AI-hallucinated citations. In a system where a wrong decision can send a person to prison, fluency is not a substitute for truth.
The Core Argument / Issue
The human-in-the-loop is an accountability rule, not a courtesy
The reason a human must author the judgment is not sentiment. It is that accountability requires an answerable author. A judge can be asked to give reasons, can be appealed against, can be held to the record and, ultimately, is subject to the discipline of the higher courts. A model can be asked for none of these. If an AISystem’s output decides a case, there is no one to hold to account for the reasoning, because there was no reasoning, only prediction. Keeping the human in the loop keeps a name on the judgment.
Where AI helps and where it must be fenced out
| Judicial function | AI role under the draft | Why the line is drawn there |
|---|---|---|
| Transcription of proceedings | Permitted, assistive | Speeds the record, human verifies accuracy |
| Translation of orders | Permitted, assistive | Widens access to justice in regional languages |
| Cause-list and scheduling | Permitted, assistive | Reduces administrative delay, no rights decided |
| Legal research and case retrieval | Permitted, with verification | Judge must confirm every citation is real |
| Judicial reasoning and drafting | Barred | Requires an accountable, answerable human author |
| Adjudication and final decision | Barred | Liberty and rights cannot rest on prediction |
Bias, hallucination and privacy are not edge cases
A model trained on past judgments can inherit and amplify the biases embedded in them, reproducing patterns of harsher treatment for particular groups under the appearance of neutral computation. It can fabricate precedent that sounds authoritative. And court records are dense with the most sensitive personal data, medical, financial, familial, which the Digital Personal Data Protection Act, 2023 requires to be processed with a lawful basis and protected against misuse. An AI tool deployed carelessly across the case management system becomes a privacy exposure at scale. Explainability, the ability to show why an output was produced, is therefore not a luxury feature but a precondition for any judicial use.
How to Think About This (Analytical Frame)
Separate the tool from the decision, and never let the second dress up as the first. Ask of any proposed AI deployment a single question: does it assist a human who remains accountable, or does it replace the human’s judgment while hiding behind the human’s signature? Transcription, translation and retrieval pass the test because a person still reads, checks and owns the result. Automated adjudication fails it because the person becomes a rubber stamp for a process they cannot interrogate. The transferable rule for your GS2 and GS3 answers: automation is legitimate where it augments an accountable human and illegitimate where it launders an unaccountable machine’s output as a human decision.
The Diagram in Words
case backlog and paperwork delay -> e-Courts digital rails built (SUVAS translation, SUPACE research) -> generative AI now able to summarise and draft -> temptation to let it decide -> draft framework draws the line -> assistive tasks permitted (transcription, translation, scheduling, research with verification) -> core judicial tasks barred (reasoning, drafting, adjudication) -> guardrails: human-in-the-loop, explainability, bias audit, DPDP Act privacy compliance -> outcome: faster access to justice without surrendering accountable human judgment
Way Forward
- Codify the assistive-adjudicatory line in binding rules. Convert the draft framework into enforceable practice directions that name the permitted assistive tasks and expressly prohibit AI from authoring reasoning or decisions, so the boundary does not blur with each new tool.
- Mandate human verification of every citation. Require that any case law surfaced by an AI research tool be independently confirmed against the official reporter before it enters an order, closing the door on hallucinated precedent.
- Audit for bias and demand explainability. Subject any model used in the justice system to bias testing on historical judgments and require that its outputs be explainable, rejecting opaque tools whose reasoning cannot be examined.
- Protect court data under the DPDP framework. Apply the Digital Personal Data Protection Act, 2023 rigorously to court records processed by AI, with data minimisation, access controls and a lawful basis for every processing operation.
PYQ Linkage and Practice
UPSC has asked about judicial delays, e-governance in the judiciary, and the challenges posed by emerging technologies. This editorial links those themes to a live governance question, how to harness AI’s efficiency without surrendering the accountability that defines a court.
Practice question: “Generative AI can assist the judicial process but must never replace the judge.” Examine the risks of deploying artificial intelligence in adjudication and evaluate the safeguards needed for its assistive use in Indian courts. (250 words, 15 marks)
Sources: The Indian Express, Supreme Court of India
Source: AI in the Courtroom: Why the Judge Must Stay in the Loop — Ujiyari.com | Free UPSC & State PCS Editorial Analysis
